TEAGER-KAISER OPERATOR BASED FILTERING
Ridha Hamila, Markku Renfors, Taneli Haverinen
Telecommunications Laboratory,
Tampere University of Technology
P.O.Box 553, FIN-33101, Tampere, Finland
ABSTRACT
In this contribution, we propose a new filtering and
denoising technique for one-dimentional signals based on
the nonlinear quadratic Teager-Kaiser operator. This
technique is a threshold ‘energy’ based approach where
outliers are first detected and then replaced by their
estimated values. The proposed technique performs
better compared to the alpha-trimmed and running mean
filter, also with moderate complexity. In particular, we
present an application of the proposed real-time filtering
approach for postprocessing the measurements of time
delay signals in a mobile positioning system highly
affected by the propagation environment.
I INTRODUCTION
Various filtering methods have been proposed in the
course of years. There are both linear and nonlinear
alternatives, and filtering is often a trade-off between
different features, such as performance in Gaussian and
non-Gaussian environments, ability to adapt in case of
nonstationary signals, edge preservation and
computational complexity. Two most typical methods are
running mean and median filters, which exploit a sliding
fixed length window. Mean filters are known to remove
well additive Gaussian noise, but they tend to soften
edges and can not fully cope with impulsive noise.
Median filters on the other hand are nonlinear, they
involve sorting of the data, and are able to eliminate
impulses at the expense of some streaking and edge
jittering. Many modifications have been developed to
find a good compromise. One of the simplest categories
is trimmed mean filters, which reject some probable
outliers before taking the mean. Ranked order and
weighted order statistic filters represent an extension of
the ordinary median filter [1].
Denoising deals basically with the same problem of
estimating the underlying signal from the noisy
observations, but data is now not needed to process on-
line or is not time-dependent at all. Therefore, noncausal
filters can be exploited, the simplest proper algorithm is
gliding window smoother, which differs from the running
window mean filter only in that the cursor is now in the
middle of the window, not in the right-hand end.
Naturally, the same procedure can be applied to the other
window-based methods as well. Denoising is not
restricted to the methods acting in time domain, the
signal is often processed in some transform domain.
Wavelet transform has recently gained research interest
in several fields of signal processing, using thresholding
Guðni Gunnarsson
Nokia Networks,
P.O.Box 361, FIN-00045, Nokia Group Finland
of (orthogonal) wavelet coefficients and taking the
inverse wavelet transform has been proposed to solve
denoising problem. However, this is a more
computationally complex approach [1].
In this paper, we propose a new filtering and denoising
approach for one-dimensional signals based on a
nonlinear quadratic operator called Teager-Kaiser (TK)
operator. As an important potential application for these
techniques, we examine improving the accuracy and
certainty of mobile phone positioning. This positioning
problem is an important and interesting research topic, as
positioning could e.g., help to locate people in need of
emergency service [2]. The location can be determined
using measurement signals from multiple Base
Transceiver Stations (BTS). However, due to the
multipath interference, a measurement signal is highly
corrupted and a proper filtering method is needed.
II TEAGER-KAISER OPERATOR
Based on Newton’s law of motion, a nonlinear quadratic
operator called Teager-Kaiser (TK) operator was first
introduced by Teager and Kaiser [3],[4] to measure the
real physical energy of a system. This nonlinear operator
differs from the common way to calculate the energy of a
discrete-time signal as the average sum of its squared
magnitudes. The energy of a generating system of a
simple oscillation signal was computed as the product of
the square of the amplitude and the frequency of the
signal. It was found that this nonlinear operator exhibits
several attractive features such as simplicity, efficiency
and ability to track instantaneouly-varying spacial
patterns. Since its introduction, several applications have
been derived for one-dimensional [5],[6], and two-
dimensional signal processing [7].
The continuous-time TK operator of a complex-valued
signal x (t ) is defined as follows [6]
EC [x(t )] = x& (t ) x&* (t ) −
1
2
[
&x&(t ) x* (t ) + x(t ) &x&* (t ) ]. (1)
When x (t ) is real, Eq. (1) reduces to the continuous-
time TK operator of a real-valued signal [3]
E R [x (t )] = x& 2 (t ) − x(t ) &x&(t ) . (2)
Similarly, by making certain combinations of the
discretized mapped derivatives, the discrete-time TK
operator for real-valued signal is given by [4]
 E D [x (n)] = E (n) = x 2 ( n) − x( n − 1) x( n + 1) . (3) TK filtering algorithm:
For n = 2 to L, do
E ( n) ← OTD 2 (n) − OTD(n − 1) × OTD(n + 1)
III E-OTD POSITIONING METHOD
Th( n) ← [Th(n − 1) × (n − 1) + β × E ( n) ] n
Mobile phones have gained tremendous popularity in the if E (n) < Th(n) then
last few years. Many kinds of services have been Tmp_OTD(n) ← OTD(n)
designed in addition to usual voice calls. The main else
characteristic of a wireless phone is its mobility; it Tmp_OTD(n) ← Filt_OTD(n − 1)
follows the user almost everywhere, and therefore one
end if
very attractive application would be a positioning service
Filt_OTD(n) ← mean[Tmp_OTD(n − win + 1 ... n)]
especially for emergency situations. A wireless phone
cannot currently be accurately located in a GSM network end do
[1],[8]. The identity of the serving Base Transceiver
Station (BTS) is called Cell Identity (CI). CI is known The adaptive threshold is estimated by β times the mean
and can be used for positioning but it results in poor of the ‘energy’ of the preceding samples. The energy-like
accuracy, which is also dependent on the size of the cell. quantity of the processed sample is then compared to the
Another, preferable way for finding a solution for this corresponding threshold value. If the ‘energy’ is greater
positioning problem would be utilizing signal time delay than the corresponding threshold, the sample value of the
or angle of arrival information [9]. In synchronous input OTD signal is replaced by the sample value of the
systems, the Observed Time Difference (OTD) previous ouput. After removing impulsive interference
positioning method calculates the time interval between through this process, the OTD signal is filtered by the
the receptions of bursts from two different BTS. This windowed mean of the preceding samples including the
results in a hyperbola that lies between the two BTS's. processed sample.
When this calculation is repeated for several neighbor
BTS's, several hyperbolas are formed, and the location of The performance of the TK filter is very much related to
the wireless phone is determined by the place where the the filter window length (win), and also to the adaptive
hyperbolas intersect. Measurement errors are parameter (β) which permits to select the best fitted
unfortunately unavoidable. One severe source of errors is threshold corresponding to the specific signal in
multi-path interference, and it leads to reduction in the question. The following table shows the MSE of the
positioning accuracy. Therefore, post-processing of the OTD filtered signal for TK filter with respect to these
measurement signal is obviously needed. In this paper, two parameters:
we used an OTD signal measured via modified software
of a NOKIA 8110 mobile phone (see Fig. 2). Here, the Table I
MSE win=5 win=10 win=15 win=20 win=25 win=30 win=35 win=40
OTD value is represented in multiple of 1/16 bit intervals β=0.3 2,17 1,99 2,18 2,57 3,06 3,65 4,33 5,11
β=0.6 2,47 1,96 1,88 2,00 2,21 2,46 2,76 3,12
which corresponds approximately to 60 meters β=0.9 2,68 1,96 1,79 1,79 1,88 2,01 2,17 2,37
β=1.2 2,95 1,98 1,70 1,65 1,68 1,74 1,84 1,97
fluctuation in distance. Notice also that the OTD signal β=1.5 3,07 1,93 1,60 1,50 1,50 1,53 1,60 1,71
β=1.8
slope varies with the MP speed. β=2.1
3,10
3,37
1,88
2,13
1,51
1,73
1,40
1,58
1,38
1,55
1,41
1,56
1,47
1,60
1,56
1,67

IV TK OPERATOR IN FILERING Notice that in case of a long window size, small values of
beta result in poor performance because the filter is too
Teager-Kaiser filtering technique is a detection based
slow to follow nonstationary variations of the signal. On
approach in which outliers are first detected using an
the other hand, when win is small, also smaller beta is
adaptive threshold (Th) based on the TK energy-like
desired to achieve good performance.
quantity, and then replaced by their estimated values. TK
filter is described by the following algorithm (formulated
Similarly, Table II provides the MSE for the alpha-
for the OTD filtering application):
trimmed mean filter [1] with respect to the window size
and the alpha parameter:
Inputs:
OTD measurement vector of length L
Table II
β adaptivity parameter MSE win=5 win=10 win=15 win=20 win=25 win=30 win=35 win=40
α=0.00 4,27 2,60 2,02 1,78 1,65 1,58 1,57 1,58
win length of the window α=0.05 4,27 2,60 2,02 1,62 1,53 1,50 1,51 1,51
α=0.1 4,27 2,60 1,81 1,62 1,48 1,46 1,47 1,50
α=0.15 4,27 2,29 1,76 1,56 1,48 1,46 1,48 1,53
Output: α=0.20 3,85 2,24 1,73 1,56 1,48 1,46 1,49 1,53
α=0.25 3,85 2,24 1,73 1,56 1,48 1,47 1,49 1,53
Filt_OTD filtered OTD signal of length L α=0.30 3,85 2,27 1,75 1,58 1,49 1,47 1,49 1,53
α=0.35 3,85 2,27 1,79 1,60 1,49 1,47 1,49 1,53
α=0.40 4,38 2,43 1,87 1,62 1,51 1,49 1,51 1,55
α=0.45 4,38 2,43 1,87 1,69 1,55 1,50 1,52 1,58
Initializations:
Tmp _ OTD = OTD
Performance comparison between the three filtering
E (1) = OTD(2)2 − OTD(1) × OTD(3) algorithms - namely Teager filter, sliding window mean,
Th = β × E (1) and alpha-trimmed mean filter - is provided for as a
function of window length in Figure 1. Only window
length is varied, other parameters are fixed to the value
that gave the minimum MSE. As we can see, for small 2540
MSE=2.166, MAXE=5.0115, Window=5, β=0.19

Teager Filter
window lengths the results are clearly beneficial for the 2520

Teager filter. As the length of the window increases, the 2500

differences get smaller, but running mean filter remains
2480
still slightly behind the other two algorithms. 0 200 400 600 800 1000 1200 1400 1600 1800 2000

2540
MSE=3.8539, MAXE=8.6667, Window=5, α=0.23

Alpha−Trimmed
At this point it is important to bear in mind that the 2520

computational complexity for the Teager filter and 2500

running mean filter is O(1), i.e., independent on the

2480
length of the window. Alpha trimmed mean involves 0 200 400 600 800 1000 1200 1400 1600 1800 2000

2540
sorting and is therefore more complex, at least MSE=4.2666, MAXE=8.06, Window=5

Running Mean
O(win*log(win)), depending on the sorting algorithm 2520

used. 2500

2480
The filter performance is also visualized for two specific 0 200 400 600 800 1000 1200
Samples (Time)
1400 1600 1800 2000

cases. Figure 2 shows the filtering results for optimum Figure 3. Filtering of OTD signal with the Teager-Kaiser
parameters, and Figure 3 shows the filtering results for filter, alpha-trimmed mean, and running mean filter,
using optimal α and β parameters. The window length is
short window length of 5, with optimal α and β. equal to 5.
30
Teager, MSE
Alpha−Trimmed, MSE
Running Mean, MSE
Teager, MAXE
V TK OPERATOR IN DENOISING
25 Alpha−Trimmed, MAXE
Running Mean, MAXE
The TK operator based denoising algorithm consists of
three steps:
3.5
20
1. Calculate the energy-like (E) and threshold (Th) for
3
each sample, and save indices of the samples for
MSE, MAXE

15 2.5 which E < Th as in Section IV.

2 2. Use linear interpolation to replace the sample-values
10 1.5
between indexed samples.
25 26 27 28 29 30
3. Use a sliding window mean to smoothen the output.

5
The algorithm used in the simulations is as follows:
Inputs:
0 OTD measurement vector of length L
0 5 10 15 20 25 30

Figure 1. MSE and maximum absolute error of Teager-

Window Length
β adaptivity parameter
Kaiser, alpha-trimmed mean, and running mean filter, in win length of the window
the OTD filtering application.
Ouput:
2550
Filt _ OTD Filtered OTD signal of length L
Input Signal

OTD Signal

2500
Initializations:
2450 Tmp _ OTD = OTD
0 200 400 600 800 1000 1200 1400 1600 1800 2000

E (1) = OTD(2) 2 − OTD (1) × OTD (3)

2540
MSE=1.4056, MAXE=3.0503, Window=25, β=1.86
Teager Filter

2520

2500
Th = β × E (1)
2480 ind = 1
0 200 400 600 800 1000 1200 1400 1600 1800 2000
2540 N =0
Alpha−Trimmed

MSE=1.4515, MAXE=3.1862, Window=30, α=0.12

2520

2500
TK denoising algorithm:
2480
0 200 400 600 800 1000 1200 1400 1600 1800 2000 for n = 2 to L, do
2540
Running Mean

E(n) ← OTD(n) 2 − OTD(n − 1) × OTD(n + 1)

MSE=1.5721, MAXE=3.0832, Window=34
2520

Th(n) ← [Th(n − 1) × (n − 1) + β × E (n)]/ n

2500

2480
0 200 400 600 800 1000 1200
Samples (Time)
1400 1600 1800 2000
if E (n) < Th(n) then
Figure 2. Filtering of OTD signal with the Teager-Kaiser ind ← [ind n]
filter, alpha-trimmed mean, and running mean filter,
using optimum parameters. N ← N +1
end if
The proposed filtering technique is used in on-line end do
applications, which require causal signal processing ind ← [ind L]
methods. On the contrary, denoising is not time-
dependent process.
 for k = 1 to N , do VI CONCLUSIONS
slope ← [Tmp _ OTD(ind (k + 1)) − Tmp _ OTD(ind ( k ))]/
In this paper we proposed a new approach for filtering
[ind (k + 1) − ind (k )] and denoising one-dimensional signals based on the
for i = ind ( k ) + 1 to ind (k + 1) − 1, do energy-like Teager operator. In particular, the filtering
Tmp _ OTD( n) ← Tmp _ OTD( n − 1) + slope approach is used for real-time postprocessing of OTD
end do signals heavily affected by multipath interference.
end do Consequently, by minimizing the variations in the OTD
Pad _ 1 ← [Tmp _ OTD(1) L Tmp _ OTD(( win − 1) / 2)] signals, we improve the positioning accuracy. The
Pad _ 2 ← [Tmp _ OTD( L + 1 − ( win − 1) / 2) L Tmp _ OTD( L)] Teager filter gives clearly better performance than the
Tmp _ OTD ← [Pad _ 1 Tmp _ OTD Pad _ 2]
other approaches when short window length is used. In
other words, the Teager filter would allow to use shorter
sum ← [Tmp _ OTD(1) L Tmp _ OTD( win)]× [1 L 1]T window length for a given MSE performance, and clearly
for t = 1 to L, do faster changes in the desired signal component could be
Filt_OTD(t) ← sum/win tolerated than with the other approaches.
sum ← sum + Tmp _ OTD(t + win + 1) − Tmp _ OTD(t )
end do ACKNOWLEDGEMENT

This work is partially supported by the Academy of Finland

As figure 4 shows, the performance of the TK denoising
under the project “Analog and Digital Signal Processing
approach is superior to the two filter types with respect Techniques for Highly integrated Tranceivers”, and Tampere
to the MSE and maximum absolute error. Figure 5 shows Graduate School in Information Science and Engineering
the denoising results for an OTD signal. (TISE).

REFERENCES
18
Smoothing Teager, MSE
Smoothing Alpha−Trimmed, MSE [1] J. Astola and P. Kuosmanen, Fundamentals of Nonlinear
16 Somoothing Running Mean, MSE
Smoothing Teager, MAXE Digital Filtering, CRC Press LLC, 1997.
Smoothing Alpha−Trimmed, MAXE
Smoothing Running Mean, MAXE
14
[2] FCC Regulation Proposal: Notice of Proposed Rule
3
12 Making, FCC Docket No. 94-237, Adopted Date: Sept.
2.5
19, 1994, Release Date: Oct. 19,1994.
MSE, MAXE

10 2

1.5
[3] J. F. Kaiser, "On a simple algorithm to calculate the
8
1
'energy' of a signal", Proc. IEEE ICASSP’90,
6
0.5 Albuquerque, New Mexico, pp. 381-384, April 1990.
0
80 85 90 95 100
4 [4] J. F. Kaiser, "On Teager's energy algorithm and its
generalization to continuous signals", Proc. 4th IEEE
2
Digital Signal Proc. Workshop, Mohonk (New Paltz),
NY, September 1990.
0
0 10 20 30 40 50 60 70 80 90 100
Window Length
Figure 4. MSE and maximum absolute error of the [5] P. Maragos, J. F. Kaiser, and T. F.Quatieri, "On amplitude
smoothing Teager-Kaiser, alpha-trimmed, and running and frequency demodulation using energy operators",
mean filter. IEEE Trans. Signal Proc., Vol. 41, pp. 1532-1550, Apr.
1993.

2540
MSE=0.41518, MAXE=1.6567, Window=100, β=0.16384
[6] R. Hamila, J. Astola, F. Alaya Cheikh, M. Gabbouj, and
Smooth. Teager

2520 M. Renfors, "Teager Energy and the Ambiguity

Function", IEEE Trans on Signal Proc., Vol. 47, No. 1,
2500
pp. 260-262, January 1999.
2480
0 200 400 600 800 1000 1200 1400 1600 1800 2000

2540 [7] F. Alaya Cheikh, R. Hamila, M. Gabbouj And J. Astola,

Smooth. α−Trimmed

MSE=0.66301, MAXE=1.895, Window=100, α=0.45098

"Impulse Noise Removal In Highly Corrupted Color
2520
Images," Proc. 1996 IEEE Int. Conf. On Image Proc.,
2500 Lausanne, Switzerland, Vol. I, pp. 997-1000, Sept.,
2480
1996.
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Smooth. Running Mean

2540
MSE=0.7973, MAXE=1.825, Window=100 [8] ETSI GSM 05.05, Digital cellular telecommunication
2520 system (Phase 2+); Radio transmission and reception,
2500
Release 1997, version 6.3.0, ETSI, 1998.
2480
0 200 400 600 800 1000 1200
Samples (Time)
1400 1600 1800 2000 [9] G. P. Yost, and S. Panchapakesan, Improvement in
Figure 5. Denoising results for optimum α and β, and Estimation of Time of Arrival (TOA) from Timing
window length 100. Advance (TA), Proc. IEEE Universal Personal
Communications (IUPC 98), Florens Italy, 1998.